Toga III 0.3 NNUE

[dkappe]

Toga III 0.3, trained on ~150m d8 positions of Toga II 4.03. Still a little shy of sf9.

https://www.patreon.com/posts/40267300

For maximum enjoyment, I recommend using my primitive NetRunner fork that eliminates the hybrid eval. Also tested on MinicNNUE.

https://github.com/dkappe/NetRunner/tree/disable-hybrid

[corres]

Toga III 0.3, trained on ~150m d8 positions of Toga II 4.03. Still a little shy of sf9.

https://www.patreon.com/posts/40267300

For maximum enjoyment, I recommend using my primitive NetRunner fork that eliminates the hybrid eval. Also tested on MinicNNUE.

https://github.com/dkappe/NetRunner/tree/disable-hybrid

Thanks!
In the old times Toga II was my favorite engine.
A question:
Using Windows 10 x64 what compiler is the most appropriate for SF + NNUE or MinicNNUE?
Mingw/clang?

[dkappe]

Ah, I only have Linux and Mac these days. Perhaps someone else can help?

[dkappe]

Just a little sampling.

[pgn]

Event "?"]
[Site "?"]
[Date "2020.08.08"]
[Round "7"]
[White "Toga3-0.3"]
[Black "SF9"]
[Result "1-0"]
[ECO "A20"]
[GameDuration "00:03:34"]
[GameEndTime "2020-08-08T04:41:24.452 CDT"]
[GameStartTime "2020-08-08T04:37:50.383 CDT"]
[Opening "English Opening"]
[PlyCount "179"]
[TimeControl "60+0.6"]

1. c4 {book} e5 {book} 2. g3 {book} c6 {book} 3. d4 {book} Bb4+ {book}
4. Bd2 {+0.07/21 2.3s} Bxd2+ {-0.15/21 3.4s} 5. Qxd2 {+0.04/21 0.95s}
e4 {-0.01/20 0.51s} 6. d5 {0.00/24 4.4s} cxd5 {-0.15/22 1.8s}
7. Qxd5 {+0.18/22 1.8s} Nf6 {+0.10/19 0.86s} 8. Qe5+ {+0.15/19 0.72s}
Qe7 {-0.17/24 7.0s} 9. Qxe7+ {+0.21/20 1.1s} Kxe7 {-0.17/1 0s}
10. Nc3 {+0.23/21 1.2s} b6 {-0.12/20 0.71s} 11. f3 {+0.31/23 2.8s}
Bb7 {-0.60/19 1.8s} 12. fxe4 {+0.27/21 0.87s} Na6 {-0.42/19 1.4s}
13. Bg2 {+0.34/21 1.4s} Rac8 {-0.66/20 2.0s} 14. b3 {+0.30/22 1.1s}
h5 {-0.56/22 6.0s} 15. Nh3 {+0.39/19 2.8s} Nc5 {-0.61/19 1.0s}
16. Nf2 {+0.42/18 0.75s} Bc6 {-0.50/19 2.0s} 17. O-O {+0.50/17 1.1s}
d6 {-0.64/20 1.9s} 18. h3 {+0.41/23 2.3s} a6 {-0.59/19 2.7s}
19. Rac1 {+0.45/20 4.5s} h4 {-0.73/21 2.6s} 20. g4 {+0.36/20 1.3s}
Rhd8 {-0.78/21 1.6s} 21. Kh2 {+0.34/22 4.6s} Re8 {-0.67/19 1.3s}
22. Rfd1 {+0.48/19 1.8s} Nfd7 {-0.80/20 2.7s} 23. Rc2 {+0.55/21 2.9s}
Ne5 {-0.90/21 2.6s} 24. Nd3 {+0.62/20 0.65s} Nexd3 {-0.97/22 2.5s}
25. exd3 {+0.63/19 1.0s} Kf8 {-0.86/21 1.3s} 26. Rf2 {+0.59/23 2.4s}
b5 {-0.79/20 0.97s} 27. Rf5 {+0.50/22 2.3s} bxc4 {-0.79/23 3.9s}
28. bxc4 {+0.52/20 0.85s} Kg8 {-0.82/22 1.8s} 29. Ne2 {+0.48/24 3.4s}
Ne6 {-0.96/22 1.9s} 30. Rd2 {+0.60/20 0.61s} Rb8 {-0.96/23 3.9s}
31. Ng1 {+0.74/23 1.8s} Nd4 {-1.29/24 7.2s} 32. Ra5 {+0.71/21 1.4s}
Rb6 {-1.26/19 0.44s} 33. Nf3 {+0.75/21 1.0s} Nxf3+ {-1.31/21 1.2s}
34. Bxf3 {+0.74/21 2.0s} Re5 {-1.22/20 0.88s} 35. Ra3 {+0.78/23 1.1s}
f6 {-1.26/21 0.71s} 36. Kg2 {+0.76/22 1.0s} Kf7 {-1.26/22 2.2s}
37. Kf2 {+0.81/23 1.0s} Bb7 {-1.31/20 0.67s} 38. Ke3 {+0.64/24 3.5s}
Rb1 {-1.16/20 0.40s} 39. Rb3 {+0.81/22 0.89s} Rxb3 {-1.00/21 1.7s}
40. axb3 {+0.70/22 2.1s} Ra5 {-0.96/20 0.64s} 41. b4 {+0.58/21 1.0s}
Ra1 {-1.18/18 0.91s} 42. Rd1 {+0.64/22 1.6s} Ra2 {-1.12/19 0.34s}
43. g5 {+0.85/21 1.3s} Rb2 {-1.14/22 2.1s} 44. gxf6 {+0.86/22 0.78s}
gxf6 {-1.10/23 0.34s} 45. Bg4 {+0.80/25 2.9s} Rxb4 {-1.11/25 0.63s}
46. Ra1 {+0.83/25 0.71s} Rb6 {-1.34/26 0.99s} 47. Ra5 {+0.67/24 2.7s}
Bc6 {-1.25/27 1.8s} 48. Kf4 {+0.66/25 0.50s} Be8 {-0.99/21 0.76s}
49. Bh5+ {+0.69/25 0.83s} Ke7 {-0.90/20 0.28s} 50. Bxe8 {+0.57/25 1.3s}
Kxe8 {-0.92/22 0.60s} 51. Rh5 {+0.57/25 0.79s} Rb1 {-0.96/21 0.56s}
52. Rxh4 {+0.64/22 0.56s} Rf1+ {-0.91/22 1.2s} 53. Ke3 {+0.83/23 0.61s}
Kd7 {-0.91/25 0.57s} 54. Rh7+ {+0.91/21 0.63s} Ke6 {-0.91/27 0.40s}
55. d4 {+0.78/23 1.1s} Re1+ {-0.57/20 0.47s} 56. Kf4 {+1.11/23 1.0s}
Rf1+ {-0.79/22 0.88s} 57. Ke3 {+0.93/22 0.78s} Re1+ {-0.60/19 0.23s}
58. Kd3 {+1.11/21 0.49s} Rd1+ {-0.47/20 0.34s} 59. Kc3 {+1.11/21 0.45s}
Rc1+ {-0.44/20 0.48s} 60. Kd2 {+1.18/23 2.1s} Rh1 {-1.04/24 2.6s}
61. h4 {+1.35/18 0.24s} a5 {-1.46/20 0.79s} 62. h5 {+1.37/18 0.37s}
Rh4 {-1.63/22 1.2s} 63. Kc3 {+1.31/19 0.55s} Rxe4 {-1.72/21 0.41s}
64. d5+ {+2.41/19 0.69s} Kf5 {-2.50/21 0.67s} 65. Rd7 {+2.45/19 0.76s}
Re3+ {-2.87/22 0.78s} 66. Kd4 {+2.55/18 0.40s} Re4+ {-2.97/22 0.60s}
67. Kd3 {+2.83/19 1.0s} Re8 {-3.14/23 0.60s} 68. Rxd6 {+2.99/17 0.36s}
Ra8 {-3.32/22 0.60s} 69. Re6 {+4.42/18 0.60s} a4 {-1.82/17 0.11s}
70. c5 {+4.77/19 0.37s} a3 {-4.70/20 1.1s} 71. Re1 {+5.74/20 0.66s}
a2 {-4.19/19 0.60s} 72. Ra1 {+6.06/19 0.68s} Ke5 {-5.98/20 0.60s}
73. Kc4 {+6.88/17 0.69s} f5 {-6.08/20 0.61s} 74. h6 {+7.14/18 0.91s}
Ra7 {-7.42/18 0.58s} 75. d6 {+7.63/15 0.42s} f4 {-8.46/18 0.60s}
76. Re1+ {+9.58/16 0.67s} Kf6 {-4.95/17 0.28s} 77. h7 {+10.35/16 0.42s}
Kg7 {-10.73/17 0.93s} 78. c6 {+11.49/14 0.43s} Kxh7 {-14.27/14 0.65s}
79. c7 {+14.75/15 0.43s} Ra8 {-25.38/14 0.56s} 80. d7 {+154.05/17 1.0s}
Ra4+ {-132.72/14 0.60s} 81. Kb5 {+154.11/17 0.47s} Re4 {-132.76/16 0.64s}
82. d8=Q {+M17/26 0.45s} a1=Q {-M16/20 0.57s} 83. Rxa1 {+M15/28 0.39s}
Re5+ {-M14/25 0.30s} 84. Kc4 {+M13/33 0.40s} Re4+ {-M12/29 0.33s}
85. Kd5 {+M11/44 0.42s} Re3 {-M10/34 0.26s} 86. Rh1+ {+M9/51 0.48s}
Kg6 {-M8/47 0.32s} 87. Qg8+ {+M7/91 0.44s} Kf6 {-M6/59 0.32s}
88. Rh6+ {+M5/245 0.35s} Ke7 {-M4/127 0.33s} 89. Qg7+ {+M3/245 0.004s}
Ke8 {-M2/1 0s} 90. c8=Q# {+M1/245 0.003s, White mates} 1-0

[/pgn]

[Leo]

Toga III 0.3, trained on ~150m d8 positions of Toga II 4.03. Still a little shy of sf9.

https://www.patreon.com/posts/40267300

For maximum enjoyment, I recommend using my primitive NetRunner fork that eliminates the hybrid eval. Also tested on MinicNNUE.

https://github.com/dkappe/NetRunner/tree/disable-hybrid

Is your goal to see how good it will get? When do you think it will max out?

[dkappe]

Toga III 0.3, trained on ~150m d8 positions of Toga II 4.03. Still a little shy of sf9.

https://www.patreon.com/posts/40267300

For maximum enjoyment, I recommend using my primitive NetRunner fork that eliminates the hybrid eval. Also tested on MinicNNUE.

https://github.com/dkappe/NetRunner/tree/disable-hybrid

Is your goal to see how good it will get? When do you think it will max out?

No idea. Maybe until it passes sf9? Sf10?

A friend of mine has been running Toga III against a large test suite. It is noticeably worse than other nets. So although it may be getting better, it may lack some fundamental positional understanding. I do like how it plays, however. Very understandable.

[mehmet123]

What' s the elo difference between Toga III 0.3 NNUE and latest standart Toga engine?

[dkappe]

What' s the elo difference between Toga III 0.3 NNUE and latest standart Toga engine?

Depends on which NNUE engine it runs in. In stockfish, 0.2 was clear of Ethereal 12.04. Suffice it to say it’s much stronger than the original it was trained from.

[Alayan]

"Toga" NNUE is Stockfish NNUE with a net trained on Toga's eval, isn't it ?

[dkappe]

"Toga" NNUE is Stockfish NNUE with a net trained on Toga's eval, isn't it ?

Unless it’s running on MinicNNUE, in which case it would be MinicNNUE Toga III by your naming scheme. When Igel and other engines support nnue nets in short order, you are free to continue contortIng yourself in naming things. I’ll save time and just refer to the net.

I’m really eager to see RofChade/Pesto with nnue support. You too?